Solar Power Ramp Events Detection Using an Optimized ...increasing levels of solar power...

NREL is a national laboratory of the U.S. Department of Energy Office of Energy Efficiency & Renewable Energy Operated by the Alliance for Sustainable Energy, LLC This report is available at no cost from the National Renewable Energy Laboratory (NREL) at www.nrel.gov/publications.

Contract No. DE-AC36-08GO28308

Solar Power Ramp Events Detection Using an Optimized Swinging Door Algorithm Preprint Mingjian Cui1, Jie Zhang2, Anthony Florita2, Bri-Mathias Hodge2, Deping Ke1, and Yuanzhang Sun1 1 Wuhan University 2 National Renewable Energy Laboratory

Presented at the ASME International Design Engineering Technical Conferences & Computers and Information in Engineering Conference Boston, Massachusetts August 2–5, 2015

Conference Paper NREL/CP-5D00-64093 August 2015

NOTICE

The submitted manuscript has been offered by an employee of the Alliance for Sustainable Energy, LLC (Alliance), a contractor of the US Government under Contract No. DE-AC36-08GO28308. Accordingly, the US Government and Alliance retain a nonexclusive royalty-free license to publish or reproduce the published form of this contribution, or allow others to do so, for US Government purposes.

This report was prepared as an account of work sponsored by an agency of the United States government. Neither the United States government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or any agency thereof.

This report is available at no cost from the National Renewable Energy Laboratory (NREL) at www.nrel.gov/publications.

Available electronically at SciTech Connect http:/www.osti.gov/scitech

Available for a processing fee to U.S. Department of Energy and its contractors, in paper, from:

U.S. Department of Energy Office of Scientific and Technical Information P.O. Box 62 Oak Ridge, TN 37831-0062 OSTI http://www.osti.gov Phone: 865.576.8401 Fax: 865.576.5728 Email: [email protected]

Available for sale to the public, in paper, from:

U.S. Department of Commerce National Technical Information Service 5301 Shawnee Road Alexandria, VA 22312 NTIS http://www.ntis.gov Phone: 800.553.6847 or 703.605.6000 Fax: 703.605.6900 Email: [email protected]

Cover Photos by Dennis Schroeder: (left to right) NREL 26173, NREL 18302, NREL 19758, NREL 29642, NREL 19795.

NREL prints on paper that contains recycled content.

http://www.osti.gov/scitechhttp://www.osti.gov/mailto:[email protected]://www.ntis.gov/mailto:[email protected]

1 This report is available at no cost from the National Renewable Energy Laboratory (NREL) at www.nrel.gov/publications.

ASME 2015 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference

IDETC/CIE 2015 August 2-5, 2015, Boston, Massachusetts, USA

DETC2015-46849

SOLAR POWER RAMP EVENTS DETECTION USING AN OPTIMIZED SWINGING DOOR ALGORITHM

Mingjian Cui1 Jie Zhang2 Anthony Florita3Wuhan University National Renewable Energy Laboratory National Renewable Energy Laboratory

Wuhan, Hubei 430072, China Golden, Colorado 80401, USA Golden, Colorado 80401, USA Email: [email protected] Email: [email protected] Email: [email protected]

Bri-Mathias Hodge4 Deping Ke5 Yuanzhang Sun6National Renewable Energy Laboratory Wuhan University Wuhan University

Golden, Colorado 80401, USA Wuhan, Hubei 430072, China Wuhan, Hubei 430072, China Email: [email protected] Email: [email protected] Email: [email protected]

ABSTRACT Solar power ramp events (SPREs) significantly influence the

integration of solar power on non-clear days and threaten the reliable and economic operation of power systems. Accurately extracting solar power ramps becomes more important with increasing levels of solar power penetrations in power systems. In this paper, we develop an optimized swinging door algorithm (OpSDA) to enhance the state of the art in SPRE detection. First, the swinging door algorithm (SDA) is utilized to segregatemeasured solar power generation into consecutive segments in a piecewise linear fashion. Then we use a dynamic programming approach to combine adjacent segments into significant ramps when the decision thresholds are met. In addition, the expected SPREs occurring in clear-sky solar power conditions areremoved. Measured solar power data from Tucson ElectricPower is used to assess the performance of the proposedmethodology. OpSDA is compared to two other ramp detection methods: the SDA and the L1-Ramp Detect with Sliding Window (L1-SW) method. The statistical results show the validity and effectiveness of the proposed method. OpSDA can significantly improve the performance of the SDA, and it can perform as well as or better than L1-SW with substantially less computation time.

Keywords: Dynamic programming, ramp forecasting, solar power ramp events, swinging door algorithm, Tucson Electric Power

NOMENCLATURE objective function to be maximized

S(i, j) score function of the time interval (i, j) R(i, j) rule sets of the time interval (i, j) RC(i, j) rule sets of the time interval (i, j) in clear-sky

time interval set of SPREs and non-SPREs time interval set of SPREs

time interval set of non-SPREs time interval of one SPRE

time interval of one non-SPRE solar power generation at the time t start time of one SPRE end time of one SPRE start time of one non-SPRE end time of one non-SPRE

m index of the mth SPRE L number of solar power data points M number of SPREs

only tunable parameter in SDA penalty parameter in L1-SW second derivative threshold in L1-SW

INTRODUCTION The increasing penetration of solar power in the United States

has presented new challenges for the reliable and economic operations of the electric grid because of the high variability and uncertainty of solar power [1-3]. Solar power ramp events

1Doctoral Candidate, School of Electrical Engineering (SEE); and Visiting

Scholar, Transmission and Grid Integration Group, National Renewable Energy Laboratory.

2Research Engineer, Transmission and Grid Integration Group, ASME Member, Corresponding author.

3Research Engineer, Transmission and Grid Integration Group, ASME Member. 4Senior Research Engineer, Transmission and Grid Integration Group. 5Lecturer, School of Electrical Engineering (SEE). 6Professor, School of Electrical Engineering (SEE).

(SPRE) are derived from large fluctuations in solar power, proposed an OpSDA to improve wind power ramp detection especially in a very short time [4]. Some of these ramp events [5] performance. are caused by changes in short-term microclimates, such as Dynamic programming is a method for solving a complex passing clouds, which are not predictable as diurnal variability. problem by breaking it down into a collection of simpler When a large SPRE occurs, it can be challenging to maintain and subproblems. It is applicable to problems exhibiting the manage a power system’s balance. During severe solar ramp properties of overlapping subproblems and optimal substructures. events, power system operators have to take measures to Sevilian and Rajagoapal [22] defined a family of scoring compensate for the significant changes in solar power. These functions with ramp event definitions and used a dynamic measures include modulating the outputs of traditional generators programming recursion technique to detect all the ramp events. (especially the thermal and hydroelectric units), using grid Boulaxis and Papadopoulos [23] utilized a dynamic ancillary services, and even curtailing or restricting the output of programming technique to solve the optimal feeder routing of solar generators without considering the economic consequences. distribution networks. Marano et al. [24] coupled a dynamic Accurately predicting and identifying SPREs reduces the programming algorithm with a mathematical model to achieve influence of solar power ramps and thereby increases the the optimal management of a compressed air energy storage dispatchability of solar power. plant.

Solar power ramp forecasting is still a relatively new research The proposed OpSDA for identifying solar power ramps is topic. However, a significant amount of work has been done in summarized as follows: solar irradiance and power forecasting, which provides useful 1) The SDA is utilized to segregate solar power data into information for solar power ramp forecasting. Zhang et al. [6] piecewise segments, and then all these piecewise segments analyzed the sensitivity of a suite of metrics to solar forecasts are merged and optimized through a dynamic programing with uniform improvement, ramp forecasting improvements, and process. To apply the dynamic programming process, the a ramp forecasting threshold. Bacher et al. [7] utilized statistical solar power signal is divided into a number of overlapping smoothing techniques and adaptive linear time series models to windows in which a ramp score function is defined to perform online forecasting of short-term solar power. Bessa et al. perform the recursion of the dynamic programming algorithm [8] presented a spatial-temporal model based on a vector and remove ramps that also occur in clear-sky power autoregressive framework to forecast solar power. It took generation. advantage of a smart grid infrastructure with smart meters and 2) Based on the OpSDA detection method, the only advanced control functions. Yang et al. [9] presented a weather- tunable parameter, , of the SDA can be ascertained as an based hybrid method for day-ahead hourly forecasting of solar optimal parameter for the online solar ramp detection. A suite power output. The method consisted of classification, training, of metrics are applied to evaluate the performance of ramp and forecasting stages. Yang et al. [10] proposed a multi- extraction with 10 different values and determine the timescale data-driven forecast model that involved the spatial and optimal value, . temporal correlations among neighboring solar sites to improve The following topics are discussed in the remainder of the the accuracy of solar power forecasting. paper: (i) the formulation of the OpSDA, which applies a

Several studies of SPRE forecasting can be found in the dynamic programming approach to the SDA; (ii) a case study of literature. Florita et al. [11] used the swinging door algorithm Tucson Electric Power (TEP), which compares the OpSDA to the (SDA) to identify variable generation ramping events from L1-Ramp Detect with Sliding Window (L1-SW) method; and (iii) historical solar power data. Hummon et al. [12] analyzed solar a further application of the OpSDA as a benchmark to tune the power ramping in the state of Gujarat in India using high- optimal value, , in the SDA. resolution solar data and found that the total magnitude of solar power ramping goes up with increased solar capacity. Hodge et OPTIMIZED SWINGING DOOR ALGORITHM al. [13] analyzed solar ramp distributions at different timescales and weather patterns. Sengupta and Keller [14] analyzed SPREs Swinging Door Algorithm (SDA) at individual locations and provided a comprehensive analysis of The SDA [15, 25] has been recently used in the literature to solar variability for a distributed generation scenario. extract ramp periods in a time series of a power signal. The SDA

In this research, a novel optimized swinging door algorithm is based on the concept of a “swinging door” with a “turning (OpSDA) is developed to detect SPREs. The proposed OpSDA point” (i.e., at time 0 with magnitude 5, as shown in Fig. 1) enhances the performance of the SDA through the introduction whenever the next point in the time series causes any of dynamic programming in the ramp detection process. The intermediate point to fall outside the area partitioned by the up SDA has been widely used in the literature for ramp forecasting. and down segment bounds. The segment bounds are defined by The SDA was originally proposed by Bristol [15] for data the door width, , which is the only tunable parameter in the compression and has been recently used in the renewable energy SDA. More detailed descriptions of the SDA can be found in [11, community. Zhang et al. [16] adopted the SDA to extract ramp 16, 17]. For instance, points A, B, and C are all inside the events from actual and forecasted wind power and evaluate the segment bounds determined by Point D within . After ramp forecasting performance of improved short-term wind segregating the power signal by the SDA, SPREs are extracted power forecasts. Makarov and Ma et al. [17-20] used the SDA to according to the user-specified definition of a significant ramp. derive three parameters for each power interval: ramping capability, ramping rate, and ramping duration. Cui et al. [21]

2


www.nrel.gov/publications

Sola

r Pow

er[p

.u.]

0.8

0.7

0 5 10 15 20 25 30 35 40 45 50 55 Samples

60 65 70

Measured data Clear-sky data Segments by SDA Up-ramps by SDA Dow n-ramps by SDA

0.6

0.5

0.4

0.3

0.2

0.1

0

FIGURE 1. SDA FOR THE EXTRACTION OF RAMPS IN THE POWER SIGNAL [11]

Fig. 2 shows one example of ramp detection results by the SDA. As shown, there are two down-ramps from Point 38 to Point 41 (points 38~40 and points 40~41); however, there should be only one significant ramp event from Point 38 to Point 41. Another case is from Point 44 to Point 47. Although there is actually one ramp event in the interval from points 44~47, the start time detected by the SDA is Point 45 rather than Point 44, which is the measured start time of this ramp event. Similar circumstances can be seen in the intervals from points 34~37 and 41~43. Moreover, the significant ramp event in the interval from points 30~33 is not detected by the SDA, because each small segment within this interval does not comply with the ramp definition. However, if the three small segments within the interval from points 30~33 were combined together into one expanded segment, it could be a significant ramp event according to the ramp definition. Therefore, it is uniquely helpful to introduce optimization techniques into the SDA to achieve more accurate ramp detection. This motivates the development of the OpSDA, which is described in the next subsection.

FIGURE 2. DETECTED SPRES BY THE SDA AND ITS CORRESPONDING SEGMENTS WITH =0.9% AND 15MINUTE RESOLUTION

OpSDA Based on a Dynamic Programming Approach The objective of the optimization is to minimize the number

of individual ramps by combining adjacent small ramps.

Therefore, adjacent segments that have the same ramping direction (e.g., up-ramps or down-ramps) can be merged into one segment. Toward this end, a dynamic programming algorithm is applied to the original segments (from the SDA). In this study, an increasing length score function, S, was designed based on the length of the interval segregated b y the SDA. The optimization seeks to maximize the length score function, which corresponds to a ramp event.

Given a solar power interval, , of all discrete time points and an objective function, , of the dynamic programming algorithm, a SPRE is detected by maximizing the objective function:

(1)s.t.

(2) (3)

where is the maximum score in the interval, , which can be computed as the maximum over (i−j) subproblems. J(i, j) is constrained by the inequality constraint in Eq. (2) and the equality constraint in Eq. (3). Eq. (2) is a super-additivity property to which the positive score value, , must conform. Eq. (3) is improved based on [22] wi th the use of a new variable,

. Except for the equality constraint with in Eq. (3), OpSDA and L1-SW share the same variables and constraints.

and are the definition of a ramp within the interval, , in measured data days and clear-sky days, respectively. Similar definitions of and in Eq. (3) have been used in the literature [26, 27] for wind power ramp detection. Generally, significant changes of solar power magnitude, direction, and duration can result in a SPRE. In this paper, we use three classic definitions of significant ramp events that were originally defined for significant wind power ramp events [16]:

(i) Significant Ramp Definition 1: The change in solar power output that is greater than 10% of the installed solar capacity.

(ii) Significant Ramp Definition 2: The change in solar power output that is greater than 10% of the installed solar capacity within a time span of 1 hour or less.

(iii) Significant Ramp Definition 3: A significant up-ramp is defined as the change in solar power output that is greater than 10% of solar capacity within a time span of 1 hour or less; and a significant down-ramp is defined as the change in solar power output that is greater than 8% of solar power capacity within a time span of 1 hour or less.

First, if and separately conform to the threshold of ramp definitions in measured data days and clear-sky days, and are assigned to be 1; otherwise,

and are assigned to be 0. Second, a comparison process is made based on Eq. (3) to remove the SPREs likewise occurring in the clear-sky days (i.e., SPREs caused by the solar diurnal variation), in which should be assigned to be 0. This comparison process is implemented based on the algorithm developed in [2 2], which is given in A lgorithm 1 w ith p seudocode. The main contribution of this research in Algorithm 1 (compared to [22]) is the introduction of a modified score

3



function, S, to remove insignificant ramps occurring both in the measured and clear-sky solar power generations. The OpSDA has the same initialization code as L1-SW in [22]. The main difference between them is how to compute maximum scores for combined segments in Algorithm 1. A flowchart is provided in Fig. 3 to illustrate the ramp detection process of the OpSDA.

FIGURE 3. THE OVERALL PROCESS OF THE OPSDA

Algorithm 1 Dynamic Programming Pseudo-Code //Length of ramp segments by the SDA [22]. L length(p)

// Initialize score of zero length segments [22]. for i = 1 L do J[i, i] 0 end for

// Compute maximum scores for combined segments. for n = 2 L do

for i = 1 L-n+1 do j i+n-1

for k = i j-1 do

// Remove clear-sky power ramps.

S(i, k) (k-i)2×R(i, k)×[1-R (i, k)] C


4

q S(i, k)+J(k, j) if q > J[i, j] then J[i, j] q

K[k, j] k end if

end for end for

end for

Based on Eqs. (1)-(3), this optimized process can proceed inductively as follows. Assuming that the number of SPREs is M and m: 1≤m

0 10200 10350 10500 10650 10800 1095Time[min]

Measured data Clear-sky data L1 trending fitting Segments of L1-SW Ramps by L1-SW Segments of OpSDA Ramps by OpSDA

0.7

0.6

0.5

0.4

Sola

r Pow

er[p

.u.]

0.3

0.2

0.1

0 11100

25

5

M easured solar power generation Clear-sky solar power generation

20

Sola

r Pow

er[M

W]

15

10

0 1000 2000 3000 4000 0

Thus, considering Eq. (2), in Eq. (6) can be transformed to:

(8) Likewise, considering Eq. (4), in Eq. (7) is:

(9)

If a solar power series is , …, , the solution is:

(10)

If a solar power series is , …, , the solution is:

(11)

Fig. 3 shows M sliding windows with the start point ( ) and the end point ( ). A set of significant ramps will be detected in each window in which with the start ( ) and end ( ). The number of significant ramps in each window may be different (e.g., the first window with and the last window with ).

In brief, all the segments (represented by the square points in Fig. 2) are first extracted by the SDA with a predefined parameter . Then all extracted segments are input into the optimization procedure (the red block shown in Fig. 3). The extracted segments are merged to yield a set of optimized significant ramp events . During this procedure, a comparison process (the blue block shown in Fig. 3) is deployed to remove the SPREs occurring in both the measured and clear-sky solar power generations. A case study of the OpSDA is shown in the next section. 0.8

Time[min] FIGURE 4. SAMPLE INTERVAL OF SOLAR POWER TIME SERIES IN TEP DATA

EXPERIMENTAL RESULTS

Data Description In this section, we present various statistics to analyze the

proposed solar ramp detection method. The solar power data collected from TEP is used for solar power ramp detection. The rated capacity of the TEP solar is 25 MW. The data includes both the measured power generation and the corresponding clear-sky power generation sampled every 15 minutes. The time interval of the TEP data spans from June 1, 2013, to October 30, 2013, including 14,564 samples. An example of the solar data time series is shown in Fig. 4.

OpSDA Ramp Event Detection Results OpSDA was compared to a ramp detection method proposed

by Sevlian and Rajagopal [22, 28], referred to as the L1-SW. The L1-SW method was developed for wind power ramp event detection, which can characterize the ramp start times, durations, and rates as well as other key features needed in the operation of a power system. In this study, the L1-SW was modified to extract SPREs. The segregation process of the L1SW method uses L1 trend fitting with a penalty parameter and the second derivative with a threshold ; whereas the OpSDA uses the SDA for segregation. The L1-SW method is capable of smoothing the noise with preprocessing in solar power and subsequently segregating the solar power into piecewise data. In this subsection, a significant ramp is defined as the change in solar power that is greater than 10% of the installed solar capacity. Note that this threshold (10%) is set relatively smaller to extract sufficient ramps to compare the L1SW to the OpSDA.

FIGURE 5. COMPARISON RESULTS OF SEGMENTS AND SIGNIFICANT RAMPS OF THE L1-SW WITH =0.5, =1×10-4 TO THE OPSDA WITH =9×10-3

Figs. 5 and 6 compare the segments approximated by the L1-SW and OpSDA methods to different parameter values. The parameter in the OpSDA is set to be 9×10-3 in both Figs. 5 and 6; the parameter in L1-SW is set to b e 0.5 and 0.02 in Figs. 5 and 6, respectively; the parameter in L1-SW is set to be 1×10-4 and 5×10-6, respectively. Significant ramps extracted by the two methods are also shown in Figs. 5 and 6.


5


0.05 0.49 OpSDA w ith ε L1-SW w ith γ = 1×10-4

L1-SW w ith γ = 1×10-5

L1-SW w ith γ = 5×10-6

Penalty Parameter λ of L1-SW

Tuna

le P

aram

eter

ε o

f OpS

DA

0.37 0.04

0.25 0.03

0.13 0.02

0.01 0.01 0 500 1000 1500 2000 2500 Runtime[s]

0.8

0.7

10200 10350 10500 10650 10800 10950 11100

Measured data Clear-sky data L1 trending fitting Segments of L1-SW Ramps by L1-SW Segments of OpSDA Ramps by OpSDA

0.6

0.5

0.4

Sola

r Pow

er[p

.u.]

0.3

0.2

0.1

0 Time[min]

Fig. 5 indicates that segments o f the OpSDA are more accurate than those of the L1-SW with larger and . For example, as shown in the intervals of the 10,600th to 10,850th minute and the 10,850th to 10,950th minute, the L1-SW method deviates from the measured solar power signal with larger and . Thus, the OpSDA has a better ramp detection performance than the L1-SW in Fig. 5. By reducing the values of and , the L1-SW method in Fig. 6 can match the measured solar power signal more precisely. However, this comes with increased computational burden.

FIGURE 6. COMPARISON RESULTS OF SEGMENTS AND SIGNIFICANT RAMPS OF THE L1-SW WITH =0.02, =5×10-6 TO THE OPSDA WITH =9×10-3

FIGURE 7. RUNTIME COMPARISON OF THE L1-SW

WITH THE PARAMETER [0.01, 0.49], =[1×10-4, 1×10-5 ,5×10-6] TO THE OPSDA WITH THE TUNABLEPARAMETER [0.01, 0.05]

Fig. 7 compares the runtime of the two methods: the OpSDA on the left y-axis and the L1-SW on the right y-axis. Although the accuracy of the L1-SW was improved in Fig. 6 with smaller and , the runtime of the L1-SW isapproximatively 40 minutes. The runtime of the OpSDA with

=9×10-3 is only 2.5 minutes. Overall, both the ramp detection performance and the runtime of the L1-SW are sensitive to the penalty parameter and the threshold . The detection performance and runtime of the OpSDA are less sensitive to the parameter . Table 1 lists the number of significant up- and down-ramps in the TEP solar power data according to the three ramp extraction methods and three definitions.

TABLE 1. NUMBER OF OBSERVED RAMPS

Methods Up-Ramps Down-Ramps

Definition 1 SDA 169 174

L1-SW 328 327OpSDA 270 296

SDA 195 201Definition 2 L1-SW 327 326

OpSDA 314 296SDA 195 222

Definition 3 L1-SW 327 388OpSDA 314 322

Table 1 shows that in all three definitions, the SDA method detects the least quantity of SPREs. Basically, there are no combinations of adjacent segments in the SDA, which makes only separate ramps conforming to specific definitions extracted. As for the L1-SW, the most quantity of SPREs is detected with smaller parameters. Consequently, this process needs the most computation. Therefore, the OpSDA detects more quantity than the SDA but with less computation than the L1-SW.

Distribution of Ramp Features A numerical statistical distribution of the three key ramp

features (ramp duration, rate, and magnitude) based on the TEP solar power data is shown in Figs. 8-10. As for ramp durations, it is observed that the most possible ramp durations detected by all three methods (SDA, OpSDA, and L1-SW) are distributed around 30 minutes with Definition 1; whereas the ramp durations are mostly distributed from 30 minutes to 60 minutes in both Definition 2 and Definition 3. Meanwhile, as for ramp rates, it is observed that the most possible ramp rates are distributed between 6 p .u./min and 13 p.u./min using all three significant ramp definitions. Moreover, this is because solar power fluctuates sharply within a short time (30 minutes or 60 minutes), as shown in Figs. 8-10.

However, as for ramp magnitudes, even though most ramp magnitudes by the three methods are distributed around 0.2 p.u., the empirical distributions derived from the L1-SW and the OpSDA are very close to each other and far from the distributions by the SDA. A sharp peak occurs when using the SDA, which indicates that there are many small solar power ramps. However, the magnitude distributions with the L1-SW and OpSDA present fat tails, which means that the magnitudes of the ramps are increased and optimized by the dynamic programming approach.



0.4

Empi

rical

Dis

trib

utio

n

0

Up-Ramps by L1-SW Down-Ramps by L1-SW Up-Ramps by OpSDA Down-Ramps by OpSDA Up-Ramps by SDA Down-Ramps by SDA

Empi

rical

Dis

trib

utio

n

30 60 90 120 150


8

6

4

2


4

3

2

1

00

Empi

rical

Dis

trib

utio

n

0.3

0.2

0.1

0 00 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0.886.7 13.3 20 26.7 33.3 40 Ramp Duration[min] Ramp Rate[p.u./min] (×10-3) Ramp Magnitude[p.u.]

(a) Ramp duration (b) Ramp rate (c) Ramp magnitude

Em

piric

al D

istr

ibut

ion

Empi

rical

Dis

trib

utio

n

0.1

0.2

0.3

0.4 Up-Ramps by L1-SW Down-Ramps by L1-SW Up-Ramps by OpSDA Down-Ramps by OpSDA Up-Ramps by SDA Down-Ramps by SDA

Empi

rical

Dis

trib

utio

n

00 30 60 90 120 150 0 6.7 13.3 20 26.7 33.3 400

2

4

6

8 Up-Ramps by L1-SW Down-Ramps by L1-SW Up-Ramps by OpSDA Down-Ramps by OpSDA Up-Ramps by SDA Down-Ramps by SDA

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80.80

1

2

3

4 Up-Ramps by L1-SW Down-Ramps by L1-SW Up-Ramps by OpSDA Down-Ramps by OpSDA Up-Ramps by SDA Down-Ramps by SDA

Ramp Duration[min] Ramp Rate[p.u./min] (×10-3) Ramp Magnitude[p.u.]

(a) Ramp duration (b) Ramp rate (c) Ramp magnitude

8

6

4

3

4

0.1

0.2

0.3

0.4 Up-Ramps by L1-SW Down-Ramps by L1-SW Up-Ramps by OpSDA Down-Ramps by OpSDA Up-Ramps by SDA Down-Ramps by SDA

Empi

rical

Dis

trib

utio

n

00 30 60 90 120 150



Empi

rical

Dis

trib

utio

n

Empi

rical

Dis

trib

utio

n

2

12

0 00 6.7 13.3 20 26.7 33.3 40 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0.88 Ramp Duration[min] Ramp Rate[p.u./min] (×10-3) Ramp Magnitude[p.u.]

(a) Ramp duration (b) Ramp rate (c) Ramp magnitude FIGURE 8. RAMP DEFINITION 1: RAMP MAGNITUDE ONLY

FIGURE 9. RAMP DEFINITION 2 : RAMP MAGNITUDE AND RAMP DURATION

FIGURE 10. RAMP DEFINITION 3 : RAMP DIRECTION, RAMP MAGNITUDE, AND RAMP DURATION

APPLICATION OF THE OPSDA FOR TUNING THE PARAMETER

The SDA has the advantage of computational and structural simplicity, which is favorable considering its robustness even with noisy data [17]. Sometimes in online applications, an SDA without any optimization is more preferable because of its inexpensive computation. Currently, this tunable parameter is normally determined through computational experiments. It would be uniquely helpful to develop an op timal strategy for adaptively selecting the best values at different time periods. Toward this end, we propose using the OpSDA as a baseline to determine the optimal parameter value, . The solving procedure is to find the that enables the ramps detected by the SDA close to that detected by the OpSDA.

Generally, a smaller value corresponds to smaller fitting errors between the SDA approximation and the measured solar power. However, an value that is too small may segregate a single significant ramp into multiple small ramps that do not satisfy the SPRE ramp definitions.

To determine the optimal parameter, , based on the

OpSDA, a suite of SPRE detection metrics were used to evaluate the performance of ramp extraction with different values. The adopted metrics include the probability of detection (POD), the critical success index (CSI), the frequency bias score (FBIAS), and the success ratio. These metrics are calculated based on a contingency table in Table 2 that provides a measure of skill for the SDA to approach the OpSDA. True positive (TP) represents the number of ramps detected by the


7


OpSDA YES

OpSDA NO Total

SDA YES TP(hits) FP(false alarm) TP+FP SDA NO FN(misses) TN FN+TN

Total TP+FN FP+TN N=TP+FP+FN+TN

10 5 3 2 1.5 1.3 1.0 10 5 3 2 1.5 1.3 1.0 11 0.9

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Def. 1: ten different ε Def. 1: εopt = 0.016



0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

06/2013 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Def. 1: ten different ε De f. 1: εopt = 0.031



0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

07/2013

0.9

Prob

abili

ty o

f Det

ectio

n Pr

obab

ility

of D

etec

tion

Prob

abili

ty o

f Det

ectio

n Pr

obab

ility

of D

etec

tion 0.8

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

0.8

0.5

0.3

0.8

0.5

0.3

Success Ratio Success Ratio (a) (b)

10 5 3 2 1.5 1.3 1.0 10 5 3 2 1.5 1.3 1.0 1

0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

1Def. 1: ten different ε Def. 1: εopt = 0.021



0.9 0.8 0.7 0.6

0.4 0.3 0.2 0.1

08/2013

0.5

0.8

0.5

0.3




0.9 0.8 0.7 0.6

0.4 0.3 0.2 0.1

09/2013

0.5

0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0.8

0.5

0.3

0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Success Ratio Success Ratio (c) (d)

SDA that are accurately detected by the OpSDA; false positive (FP) is the number of ramps detected by the SDA that are not detected by the OpSDA; false negative (FN) represents the number of ramps detected by the OpSDA that are not extracted by the SDA; true negative (TN) is the number of nonoccurring events for both the SDA and the OpSDA; and N is the total number of SP REs.

TABLE 2. CONTINGENCY TABLE FOR THE SDA AND OPSDA

Categorical statistics provide measures of accuracy and skill for forecasting in many applications, such as ramps in power, temperature, or rainfall. Based on the contingency table, a suite of metrics can be derived for optimal performance

evaluation. The equations of POD, CSI, FBIAS, and FAR can be found in [ 16] and [29].

After calculating all of the metrics (POD, CSI, FBIAS, and FAR), the optimal value is determined by the largest POD, which can be visualized on the performance diagram shown in Fig. 11. A performance diagram is used to understand the evolution of the SDA wi th different values. For the performance diagram shown in Fig. 11, (i) the left axis represents the value of POD; (ii) the bottom axis represents the success ratio; (iii) the diagonal dashed lines represent FBIAS; and (iv) the dashed curves show CSI. Fig. 11 shows a performance space of 10 values by utilizing each ramp definition. These 10 parameters are derived from 0.006 to 0.096 uniformly. To find the optimal parameter, , the points in Fig. 11 move toward the top right corner of the performance diagram. Fig. 11 shows the optimal parameters for four months (June, J uly, August, and September 2013) using 15-minute resolution data based on three significant solar power ramp definitions.

FIGURE 11. RAMP DETECTION PERFORMANCE WITH THE OPTIMAL PARAMETERS. THE SUBFIGURES (A), (B), (C), AND ( D) SHOW OPTIMAL PARAMETERS DEFINED BY THREE RAMP DEFINITIONS IN JUNE, JULY, AUGUST, AND SEPTEMBER 2013, RESPECTIVELY


8


It is observed in Fig. 11 that in June, July, and August, the [5] Cui, M., Ke, D., Sun, Y., Gan, D., Zhang, J., and Hodge, B.optimal parameters have the same values under the three M., 2015, “Wind Power Ramp Event Forecasting Using a definitions: 0.016, 0.031, and 0.021, respectively. This means Stochastic Scenario Generation Method,” IEEE Transactions

on Sustainable Energy, DOI: 10.1109/TSTE.2014.2386870. that the optimal is not sensitive to the ramp definitions in [6] Zhang, J., Florita, A., Hodge, B.-M., Lu, S., Hamann, H. F., these three months; the determined optimal values are robust. Banunarayanan, V., and Brockway, A., 2015, “A Suite of Meanwhile, when using the optimal parameter, , both the Metrics for Assessing the Performance of Solar Power POD and SR values are larger than 0.7 in all three months; the Forecasting,” Solar Energy, 111, pp. 157-175. FBAIS values are close to 1; and the CSI values are larger than [7] Bacher, P., Madsen, H., and Nielsen, H. A., 2009, “Online 0.6. This shows that the ramp detection results obtained by the Short-Term Solar Power Forecasting,” Solar Energy, 83(10), SDA with can successfully approach the results obtained pp. 1772-1783.

[8] Bessa, R. J., Trindade A., and Miranda, V., 2015, “Spatial-by the OpSDA. In September, Definition 3 has a smaller Temporal Solar Power Forecasting for Smart Grids,” IEEE 0.011 than Definition 1 and Definition 2, in which = 0.046. Transactions on Industrial Informatics, 11(1), pp. 232-241. This can guide power system operators to make decisions about [9] Yang, H. T., Huang, C. M., Huang, Y. C., and Pai, Y. S., 2014, how to best capture solar power ramps. “A Weather-Based Hybrid Method for 1-Day Ahead Hourly

Forecasting of PV Power Output,” IEEE Transactions on CONCLUSION Sustainable Energy, 5(3), pp. 917-926.

An OpSDA was developed in this paper for s olar power [10] Yang, C., Thatte, A. A., and Xie, L., 2015, “Multitime-Scale Data-Driven Spatio-Temporal Forecast of Photovoltaic ramp detection. The developed OpSDA was compared to the Generation,” IEEE Transaction on Sustainable Energy, 6(1), SDA and L1-SW methods in a case study of solar power ramp pp. 104-112.

extraction from TEP data. The results showed that the OpSDA [11] Florita, A., Hodge, B.-M., and Orwig, K., 2013, “Identifying successfully identified SPREs and performed significantly Wind and Solar Ramping Events,” IEEE 5th Green better than the SDA. The OpSDA provided equal-to-better Technologies Conference, Denver, CO. performance than the L1-SW method with much less [12] Hummon, M., Cochran, J., Weekley, A., Lopez, A., Zhang, J., computational time. The developed OpSDA was also used as a Stoltenberg, B., Parsons, B., Batra, P., Mehta, B., and Patel, D., baseline to determine the optimal values of the tunable 2014, “Variability of Photovoltaic Power in the State of parameter in the SDA. Gujarat Using High Resolution Solar Data,” Technical Report,

No. NREL/TP-7A40-60991, National Renewable Energy In future work, more data at multiple spatial and temporal Laboratory, Golden, CO. resolutions will be analyzed using this ramp detection method [13] Hodge, B.-M., Hummon, M., and Orwig, K., 2011, “Solar OpSDA for further performance evaluation. In addition, some Ramping Distributions over Multiple Timescales and Weather sufficiently small non-SPREs (termed “bumps”)⎯which have Patterns,” 1st International Workshop on Integration of Solar very short time durations, small ramp magnitudes, and opposite Power into Power Systems, Aarhus, Denmark. ramping directions⎯can abruptly interrupt the iteration of [14] Sengupta, M., and Keller, J., 2012, “PV Ramping in a dynamic programming. Future work will merge bumps into the Distributed Generation Environment: A Study Using Solar adjacent segments in the process of dynamic programming for Measurements,” 38th IEEE Photovoltaic Specialists

Conference, Austin, TX. better ramp detection. [15] Bristol, E. H., 1990, “Swinging Door Trending: Adaptive Trend Recording?” ISA National Conference Proceedings.

ACKNOWLEDGMENTS [16] Zhang, J., Florita, A., Hodge, B.-M., and Freedman, J., 2014, This work was supported by the U.S. Department of Energy “Ramp Forecasting Performance From Improved Short-Term

under Contract No. DE-AC36-08-GO28308 with the National Wind Power Forecasting,” ASME International Design Renewable Energy Laboratory and by the National Basic Engineering Technical Conference & Computers and Research Program of China (2012CB215101). Information in Engineering Conference, Buffalo, NY.

[17] Makarov, Y. V., Loutan, C., Ma, J., and Mello, P., 2009, “Operational Impacts of Wind Generation on California Power REFERENCES Systems,” IEEE Transactions on Power System, 24(2), pp. [1] Yan, R., and Saha, T. K., 2012, “Voltage Variation Sensitivity 1039-1050. Analysis for Unbalanced Distribution Networks due to [18] Ma, J., Makarov, Y. V., Loutan, C., and Xie, Z., 2011, “Impact Photovoltaic Power Fluctuations,” IEEE Transactions on of Wind and Solar Generation on the California ISO’s Intra-Power Systems, 27(2), pp. 1078-1089. Hour Balancing Needs,” IEEE Power & Energy Society [2] Lave, M., and Kleissl, J., 2010, “Solar Variability of Four Sites General Meeting, San Diego, CA. Across the State of Colorado,” Renewable Energy, 35(12), pp. [19] Ma, J., Lu, S., Etingov, P. V., and Makarov, Y. V., 2012, 2867-2873. “Evaluating the Impact of Solar Generation on Balancing [3] Yan, R., and Saha, T. K., 2011, “Investigation of Voltage Requirements in Southern Nevada System,” IEEE Power & Variations in Unbalanced Distribution Systems due to High Energy Society General Meeting, San Diego, CA. Photovoltaic Penetrations,” IEEE Power & Energy Society [20] Ma, J., Lu, S., Hafen, R. P., Etingov, P. V., Makarov, Y. V., General Meeting, San Diego, CA. and Chadliev, V., 2012, “The Impact of Solar Photovoltaic [4] Godfrey, T., Mullen, S., Dugan, R. C., Rodine, C., Griffith, D. Generation on Balancing Requirements in the Southern Nevada W., and Golmie, N., 2010, “Modeling Smart Grid Applications System,” IEEE Transmission and Distribution Conference and with Co-Simulation,” 1st IEEE International Conference on Exposition, Orlando, FL. Smart Grid Communications, Gaithersburg, MD.

9




[21] Cui, M., Zhang, J., Florita, A. R., Hodge, B.-M., Ke, D., and [25] Barr, D. C., 1994, “The Use of a Data Historian to Extend Sun, Y., 2015, “An Optimized Swinging Door Algorithm for Plant Life,” International Conference on Life Management of Wind Power Ramp Events Detection,” IEEE Power & Energy Power Plants, Edinburgh, U.K. Society General Meeting, Denver, CO. [26] Kamath, C., 2010, “Understanding Wind Ramp Events

[22] Sevlian, R., and Rajagopal, R., 2013, “Detection and Statistics Through Analysis of Historical Data,” IEEE Power & Energy of Wind Power Ramps,” IEEE Transactions on P ower System, Society Transmission and Distribution Conference and 28(4), pp. 3610-3620. Exposition, New Orleans, LA.

[23] Boulaxis, N. G., and Papadopoulos, M. P., 2002, “Optimal [27] Potter, C. W., Grimit, E., and Nijssen, B., 2009, “Potential Feeder Routing in Distribution System Planning Using Benefits of a Dedicated Probabilistic Rapid Ramp Event Dynamic Programming Technique and GIS Facilities,” IEEE Forecast Tool,” IEEE Power & Energy Society Power Systems Transactions on Power Delivery, 17(1), pp. 242-247. Conference and Exposition, Seattle, WA.

[24] Marano, V., Rizzo, G., and Tiano, F. A., 2012, “Application of [28] Sevlian, R., and Rajagopal, R., 2012, “Wind Power Ramps: Dynamic Programming to the Optimal Management of a Detection and Statistics,” IEEE Power & Energy Society Hybrid Power Plant with Wind Turbines, Photovoltaic Panels General Meeting, San Diego, CA. and Compressed Air Energy Storage,” Applied Energy, pp. [29] Roebber, P. J., 2009, “Visualizing Multiple Measures of 849-859. Forecast Quality,” Weather and Forecasting, 24(2), pp. 601

608.


64093 AAs web.pdfAbstract0F(INTRODUCTIONoptimized swinging door algorithmSwinging Door Algorithm (SDA)OpSDA Based on a Dynamic Programming Approach

EXPERIMENTAL RESULTSData DescriptionOpSDA Ramp Event Detection ResultsDistribution of Ramp Features

Application of the OpSDA for Tuning the ParameterConclusionAcknowledgmentsReferences

Solar Power Ramp Events Detection Using an Optimized ...increasing levels of solar power...

Documents

Transcript of Solar Power Ramp Events Detection Using an Optimized ...increasing levels of solar power...